재난 상황에서의 생성형 언어 모델에 대한 백도어 공격 시나리오와 실험적 분석

초록

With the remarkable advancement of artificial intelligence, AI models have begun to be used across various fields, while simultaneouslyrevealing security vulnerabilities within AI itself. This paper assumes a situation in which a backdoor is inserted into a generative modelused in disaster situations through data poisoning and transfer learning. By means of this backdoor, the model produces normal outputsfor general inputs, yet deliberately yields manipulated outputs for inputs containing a specific trigger, thereby potentially causing socialconfusion. We experimentally demonstrate this possibility and emphasize its associated risks.

키워드